linux/lib/group_cpus.c
Ming Lei 0263f92fad lib/group_cpus.c: avoid acquiring cpu hotplug lock in group_cpus_evenly
group_cpus_evenly() could be part of storage driver's error handler, such
as nvme driver, when may happen during CPU hotplug, in which storage queue
has to drain its pending IOs because all CPUs associated with the queue
are offline and the queue is becoming inactive.  And handling IO needs
error handler to provide forward progress.

Then deadlock is caused:

1) inside CPU hotplug handler, CPU hotplug lock is held, and blk-mq's
   handler is waiting for inflight IO

2) error handler is waiting for CPU hotplug lock

3) inflight IO can't be completed in blk-mq's CPU hotplug handler
   because error handling can't provide forward progress.

Solve the deadlock by not holding CPU hotplug lock in group_cpus_evenly(),
in which two stage spreads are taken: 1) the 1st stage is over all present
CPUs; 2) the end stage is over all other CPUs.

Turns out the two stage spread just needs consistent 'cpu_present_mask',
and remove the CPU hotplug lock by storing it into one local cache.  This
way doesn't change correctness, because all CPUs are still covered.

Link: https://lkml.kernel.org/r/20231120083559.285174-1-ming.lei@redhat.com
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reported-by: Yi Zhang <yi.zhang@redhat.com>
Reported-by: Guangwu Zhang <guazhang@redhat.com>
Tested-by: Guangwu Zhang <guazhang@redhat.com>
Reviewed-by: Chengming Zhou <zhouchengming@bytedance.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Cc: Keith Busch <kbusch@kernel.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-12-06 16:12:46 -08:00

440 lines
11 KiB
C

// SPDX-License-Identifier: GPL-2.0
/*
* Copyright (C) 2016 Thomas Gleixner.
* Copyright (C) 2016-2017 Christoph Hellwig.
*/
#include <linux/kernel.h>
#include <linux/slab.h>
#include <linux/cpu.h>
#include <linux/sort.h>
#include <linux/group_cpus.h>
#ifdef CONFIG_SMP
static void grp_spread_init_one(struct cpumask *irqmsk, struct cpumask *nmsk,
unsigned int cpus_per_grp)
{
const struct cpumask *siblmsk;
int cpu, sibl;
for ( ; cpus_per_grp > 0; ) {
cpu = cpumask_first(nmsk);
/* Should not happen, but I'm too lazy to think about it */
if (cpu >= nr_cpu_ids)
return;
cpumask_clear_cpu(cpu, nmsk);
cpumask_set_cpu(cpu, irqmsk);
cpus_per_grp--;
/* If the cpu has siblings, use them first */
siblmsk = topology_sibling_cpumask(cpu);
for (sibl = -1; cpus_per_grp > 0; ) {
sibl = cpumask_next(sibl, siblmsk);
if (sibl >= nr_cpu_ids)
break;
if (!cpumask_test_and_clear_cpu(sibl, nmsk))
continue;
cpumask_set_cpu(sibl, irqmsk);
cpus_per_grp--;
}
}
}
static cpumask_var_t *alloc_node_to_cpumask(void)
{
cpumask_var_t *masks;
int node;
masks = kcalloc(nr_node_ids, sizeof(cpumask_var_t), GFP_KERNEL);
if (!masks)
return NULL;
for (node = 0; node < nr_node_ids; node++) {
if (!zalloc_cpumask_var(&masks[node], GFP_KERNEL))
goto out_unwind;
}
return masks;
out_unwind:
while (--node >= 0)
free_cpumask_var(masks[node]);
kfree(masks);
return NULL;
}
static void free_node_to_cpumask(cpumask_var_t *masks)
{
int node;
for (node = 0; node < nr_node_ids; node++)
free_cpumask_var(masks[node]);
kfree(masks);
}
static void build_node_to_cpumask(cpumask_var_t *masks)
{
int cpu;
for_each_possible_cpu(cpu)
cpumask_set_cpu(cpu, masks[cpu_to_node(cpu)]);
}
static int get_nodes_in_cpumask(cpumask_var_t *node_to_cpumask,
const struct cpumask *mask, nodemask_t *nodemsk)
{
int n, nodes = 0;
/* Calculate the number of nodes in the supplied affinity mask */
for_each_node(n) {
if (cpumask_intersects(mask, node_to_cpumask[n])) {
node_set(n, *nodemsk);
nodes++;
}
}
return nodes;
}
struct node_groups {
unsigned id;
union {
unsigned ngroups;
unsigned ncpus;
};
};
static int ncpus_cmp_func(const void *l, const void *r)
{
const struct node_groups *ln = l;
const struct node_groups *rn = r;
return ln->ncpus - rn->ncpus;
}
/*
* Allocate group number for each node, so that for each node:
*
* 1) the allocated number is >= 1
*
* 2) the allocated number is <= active CPU number of this node
*
* The actual allocated total groups may be less than @numgrps when
* active total CPU number is less than @numgrps.
*
* Active CPUs means the CPUs in '@cpu_mask AND @node_to_cpumask[]'
* for each node.
*/
static void alloc_nodes_groups(unsigned int numgrps,
cpumask_var_t *node_to_cpumask,
const struct cpumask *cpu_mask,
const nodemask_t nodemsk,
struct cpumask *nmsk,
struct node_groups *node_groups)
{
unsigned n, remaining_ncpus = 0;
for (n = 0; n < nr_node_ids; n++) {
node_groups[n].id = n;
node_groups[n].ncpus = UINT_MAX;
}
for_each_node_mask(n, nodemsk) {
unsigned ncpus;
cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
ncpus = cpumask_weight(nmsk);
if (!ncpus)
continue;
remaining_ncpus += ncpus;
node_groups[n].ncpus = ncpus;
}
numgrps = min_t(unsigned, remaining_ncpus, numgrps);
sort(node_groups, nr_node_ids, sizeof(node_groups[0]),
ncpus_cmp_func, NULL);
/*
* Allocate groups for each node according to the ratio of this
* node's nr_cpus to remaining un-assigned ncpus. 'numgrps' is
* bigger than number of active numa nodes. Always start the
* allocation from the node with minimized nr_cpus.
*
* This way guarantees that each active node gets allocated at
* least one group, and the theory is simple: over-allocation
* is only done when this node is assigned by one group, so
* other nodes will be allocated >= 1 groups, since 'numgrps' is
* bigger than number of numa nodes.
*
* One perfect invariant is that number of allocated groups for
* each node is <= CPU count of this node:
*
* 1) suppose there are two nodes: A and B
* ncpu(X) is CPU count of node X
* grps(X) is the group count allocated to node X via this
* algorithm
*
* ncpu(A) <= ncpu(B)
* ncpu(A) + ncpu(B) = N
* grps(A) + grps(B) = G
*
* grps(A) = max(1, round_down(G * ncpu(A) / N))
* grps(B) = G - grps(A)
*
* both N and G are integer, and 2 <= G <= N, suppose
* G = N - delta, and 0 <= delta <= N - 2
*
* 2) obviously grps(A) <= ncpu(A) because:
*
* if grps(A) is 1, then grps(A) <= ncpu(A) given
* ncpu(A) >= 1
*
* otherwise,
* grps(A) <= G * ncpu(A) / N <= ncpu(A), given G <= N
*
* 3) prove how grps(B) <= ncpu(B):
*
* if round_down(G * ncpu(A) / N) == 0, vecs(B) won't be
* over-allocated, so grps(B) <= ncpu(B),
*
* otherwise:
*
* grps(A) =
* round_down(G * ncpu(A) / N) =
* round_down((N - delta) * ncpu(A) / N) =
* round_down((N * ncpu(A) - delta * ncpu(A)) / N) >=
* round_down((N * ncpu(A) - delta * N) / N) =
* cpu(A) - delta
*
* then:
*
* grps(A) - G >= ncpu(A) - delta - G
* =>
* G - grps(A) <= G + delta - ncpu(A)
* =>
* grps(B) <= N - ncpu(A)
* =>
* grps(B) <= cpu(B)
*
* For nodes >= 3, it can be thought as one node and another big
* node given that is exactly what this algorithm is implemented,
* and we always re-calculate 'remaining_ncpus' & 'numgrps', and
* finally for each node X: grps(X) <= ncpu(X).
*
*/
for (n = 0; n < nr_node_ids; n++) {
unsigned ngroups, ncpus;
if (node_groups[n].ncpus == UINT_MAX)
continue;
WARN_ON_ONCE(numgrps == 0);
ncpus = node_groups[n].ncpus;
ngroups = max_t(unsigned, 1,
numgrps * ncpus / remaining_ncpus);
WARN_ON_ONCE(ngroups > ncpus);
node_groups[n].ngroups = ngroups;
remaining_ncpus -= ncpus;
numgrps -= ngroups;
}
}
static int __group_cpus_evenly(unsigned int startgrp, unsigned int numgrps,
cpumask_var_t *node_to_cpumask,
const struct cpumask *cpu_mask,
struct cpumask *nmsk, struct cpumask *masks)
{
unsigned int i, n, nodes, cpus_per_grp, extra_grps, done = 0;
unsigned int last_grp = numgrps;
unsigned int curgrp = startgrp;
nodemask_t nodemsk = NODE_MASK_NONE;
struct node_groups *node_groups;
if (cpumask_empty(cpu_mask))
return 0;
nodes = get_nodes_in_cpumask(node_to_cpumask, cpu_mask, &nodemsk);
/*
* If the number of nodes in the mask is greater than or equal the
* number of groups we just spread the groups across the nodes.
*/
if (numgrps <= nodes) {
for_each_node_mask(n, nodemsk) {
/* Ensure that only CPUs which are in both masks are set */
cpumask_and(nmsk, cpu_mask, node_to_cpumask[n]);
cpumask_or(&masks[curgrp], &masks[curgrp], nmsk);
if (++curgrp == last_grp)
curgrp = 0;
}
return numgrps;
}
node_groups = kcalloc(nr_node_ids,
sizeof(struct node_groups),
GFP_KERNEL);
if (!node_groups)
return -ENOMEM;
/* allocate group number for each node */
alloc_nodes_groups(numgrps, node_to_cpumask, cpu_mask,
nodemsk, nmsk, node_groups);
for (i = 0; i < nr_node_ids; i++) {
unsigned int ncpus, v;
struct node_groups *nv = &node_groups[i];
if (nv->ngroups == UINT_MAX)
continue;
/* Get the cpus on this node which are in the mask */
cpumask_and(nmsk, cpu_mask, node_to_cpumask[nv->id]);
ncpus = cpumask_weight(nmsk);
if (!ncpus)
continue;
WARN_ON_ONCE(nv->ngroups > ncpus);
/* Account for rounding errors */
extra_grps = ncpus - nv->ngroups * (ncpus / nv->ngroups);
/* Spread allocated groups on CPUs of the current node */
for (v = 0; v < nv->ngroups; v++, curgrp++) {
cpus_per_grp = ncpus / nv->ngroups;
/* Account for extra groups to compensate rounding errors */
if (extra_grps) {
cpus_per_grp++;
--extra_grps;
}
/*
* wrapping has to be considered given 'startgrp'
* may start anywhere
*/
if (curgrp >= last_grp)
curgrp = 0;
grp_spread_init_one(&masks[curgrp], nmsk,
cpus_per_grp);
}
done += nv->ngroups;
}
kfree(node_groups);
return done;
}
/**
* group_cpus_evenly - Group all CPUs evenly per NUMA/CPU locality
* @numgrps: number of groups
*
* Return: cpumask array if successful, NULL otherwise. And each element
* includes CPUs assigned to this group
*
* Try to put close CPUs from viewpoint of CPU and NUMA locality into
* same group, and run two-stage grouping:
* 1) allocate present CPUs on these groups evenly first
* 2) allocate other possible CPUs on these groups evenly
*
* We guarantee in the resulted grouping that all CPUs are covered, and
* no same CPU is assigned to multiple groups
*/
struct cpumask *group_cpus_evenly(unsigned int numgrps)
{
unsigned int curgrp = 0, nr_present = 0, nr_others = 0;
cpumask_var_t *node_to_cpumask;
cpumask_var_t nmsk, npresmsk;
int ret = -ENOMEM;
struct cpumask *masks = NULL;
if (!zalloc_cpumask_var(&nmsk, GFP_KERNEL))
return NULL;
if (!zalloc_cpumask_var(&npresmsk, GFP_KERNEL))
goto fail_nmsk;
node_to_cpumask = alloc_node_to_cpumask();
if (!node_to_cpumask)
goto fail_npresmsk;
masks = kcalloc(numgrps, sizeof(*masks), GFP_KERNEL);
if (!masks)
goto fail_node_to_cpumask;
build_node_to_cpumask(node_to_cpumask);
/*
* Make a local cache of 'cpu_present_mask', so the two stages
* spread can observe consistent 'cpu_present_mask' without holding
* cpu hotplug lock, then we can reduce deadlock risk with cpu
* hotplug code.
*
* Here CPU hotplug may happen when reading `cpu_present_mask`, and
* we can live with the case because it only affects that hotplug
* CPU is handled in the 1st or 2nd stage, and either way is correct
* from API user viewpoint since 2-stage spread is sort of
* optimization.
*/
cpumask_copy(npresmsk, data_race(cpu_present_mask));
/* grouping present CPUs first */
ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
npresmsk, nmsk, masks);
if (ret < 0)
goto fail_build_affinity;
nr_present = ret;
/*
* Allocate non present CPUs starting from the next group to be
* handled. If the grouping of present CPUs already exhausted the
* group space, assign the non present CPUs to the already
* allocated out groups.
*/
if (nr_present >= numgrps)
curgrp = 0;
else
curgrp = nr_present;
cpumask_andnot(npresmsk, cpu_possible_mask, npresmsk);
ret = __group_cpus_evenly(curgrp, numgrps, node_to_cpumask,
npresmsk, nmsk, masks);
if (ret >= 0)
nr_others = ret;
fail_build_affinity:
if (ret >= 0)
WARN_ON(nr_present + nr_others < numgrps);
fail_node_to_cpumask:
free_node_to_cpumask(node_to_cpumask);
fail_npresmsk:
free_cpumask_var(npresmsk);
fail_nmsk:
free_cpumask_var(nmsk);
if (ret < 0) {
kfree(masks);
return NULL;
}
return masks;
}
#else /* CONFIG_SMP */
struct cpumask *group_cpus_evenly(unsigned int numgrps)
{
struct cpumask *masks = kcalloc(numgrps, sizeof(*masks), GFP_KERNEL);
if (!masks)
return NULL;
/* assign all CPUs(cpu 0) to the 1st group only */
cpumask_copy(&masks[0], cpu_possible_mask);
return masks;
}
#endif /* CONFIG_SMP */
EXPORT_SYMBOL_GPL(group_cpus_evenly);